Hacker News Reader

Comment by mrb | Hacker News Reader

It is surprising the paper makes no comparisons with GPUs. So I will do it.

For starters it looks like they are talking about integer operations (I only skimmed the paper and it mentions an ALU, not an FPU), whereas my GPU numbers below are single precision floating point numbers. So it is apples vs oranges.

So, a modern 14-16nm GPU like the Tesla P100 or RX 480 does about 5 to 10 trillion ops/sec at 200-300 W, and approximately 30 pJ/op. So GPUs are about 5x less power efficient. The paper authors did a good job. However GPUs are not optimized for absolute power efficiency, but mostly for performance per $.

View on HN

Another big difference is that most GPU architectures are multi-lane SIMD (so single instructions acting on multiple data but multiple sets of those) whereas the linked architure is MIMD.

In simpler terms: these processors all execute independent code whereas a GPU tends to have multiple cores but a (sometimes much) smaller number of threads of execution.

Nvidia switched to MIMD in their Fermi architecture. I assume that they are the reason why jacquesm says most.

Eh, not really. 32 'threads' share a program counter on Fermi; it's still actually pretty heavily SIMD.

Also important is the memory bandwidth. You can do all the TFLOPs you want, but if you can't access the data fast enough, then there's a bottleneck.

(Dynamic) Power consumption goes up with square of voltage, so a comparison with a gpu's max power efficiency point (almost certainly underclocked and undervolted) could change the comparison significantly.

Power consumption goes up with square of voltage

This needs to become common knowledge amongst nerds and tech types.

It was more widely known, when every geek overclocked; no more.

I was underclocking when everyone else was overclocking. Because I wanted quiet!

You are very right. I should have mentioned it in my comment. Back when I was mining bitcoins on AMD GPUs, undervolting and underclocking would easily double or triple the power efficiency.

Cursory Google search shows Pascal GPUs at 10pJ/op for double precision (these guys did int, which should be 10x faster)[1,2]

[1] http://wccftech.com/nvidia-pascal-volta-gpus-sc15/

[2] http://prism.sejong.ac.kr/download/PRISM3_paper/PRISM_Ajayku...

Reply on news.ycombinator.com